Smart adaptive learning and optimized feature clustering for enhanced image retrieval

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Retrieval

Background:

• Content-Based Image Retrieval (CBIR) faces challenges with vast multimedia data.
• Accurate feature dissimilarities are crucial for effective CBIR.
• Existing methods struggle with premature convergence and redundant activations.

Purpose of the Study:

• To propose a novel approach, SEGJO-EDCNN, for enhanced CBIR.
• To improve feature clustering and matching accuracy in image retrieval.
• To address limitations of existing CBIR techniques.

Main Methods:

• Developed SEGJO (Scaling Factor and Elite Opposition Learning-based Golden Jackal Optimization) for feature clustering.
• Integrated Scaling Factor (SF) and Elite Opposition Learning (EOL) to enhance search and prevent premature convergence.
• Utilized Local Binary Pattern, Zernike Moments, and Color Moments for feature extraction.
• Incorporated an Entropy-based Divergence (ED) function within a Convolutional Neural Network (CNN) named EDCNN for improved matching.

Main Results:

• SEGJO-EDCNN demonstrated superior performance on Corel 5K and Oxford Flower datasets.
• Achieved a mean average precision (MAP) of 97.595% on the Corel 5K dataset, outperforming ELNDP and DNN-SAR.
• Attained a MAP of 99.239% on the Oxford Flower dataset, surpassing SVM-CBIR.

Conclusions:

• The proposed SEGJO-EDCNN method significantly enhances CBIR performance.
• The integration of SF, EOL, and EDCNN effectively improves feature clustering and retrieval accuracy.
• SEGJO-EDCNN offers a robust solution for the growing challenges in large-scale image retrieval.